Pose ResNet: 3D Human Pose Estimation Based on Self-Supervision.

Area of Science:

• Computer Vision
• Machine Learning
• Human-Computer Interaction

Background:

• Accurate 3D human pose estimation is crucial for applications like motion recognition and autonomous driving.
• Acquiring 3D ground truth data for training pose estimation models is challenging and resource-intensive.

Purpose of the Study:

• To propose a novel self-supervised 3D human pose estimation model (Pose ResNet) that utilizes single 2D images.
• To overcome the limitations of requiring 3D ground truth labels for training.

Main Methods:

• Utilized ResNet50 as the backbone for feature extraction.
• Incorporated Convolutional Block Attention Module (CBAM) for pixel refinement and Waterfall Atrous Spatial Pooling (WASP) for multi-scale feature capture.
• Employed a self-supervised training strategy using epipolar geometry transformation to generate 3D labels from 2D images, alongside transfer learning and synthetic occlusion.

Main Results:

• Achieved a Mean Per Joint Position Error (MPJPE) of 74.6 mm without using 3D ground truth data.
• Demonstrated accurate 3D human pose estimation from single 2D images.

Conclusions:

• The proposed Pose ResNet model effectively performs self-supervised 3D human pose estimation from 2D images.
• The method offers a viable solution for scenarios where 3D ground truth data is unavailable, showing competitive performance.